I expected to hate Claude Design. I didn't.

I've been "AI-native" in my design practice for a while now. Most of my production work at Sourceful starts in Claude Code rather than Figma, building components directly against the design system, skipping the static mockup to dev handoff loop entirely. So when Anthropic shipped Claude Design, I was sceptical. Another AI design tool promising to replace the canvas? I've seen that pitch before.

I spent today using it properly, on real work, and I have to eat a small slice of that scepticism.

The job

We've just rebranded at Sourceful, which means every social media lock-up needs rebuilding. Posts, partnership announcements, product launches, blog content, the full set. Across all the main networks. Static and animated. It's exactly the kind of volume work that eats a week of a designer's time for diminishing creative returns.

A good test, in other words.

The workflow

1. Feeding it the design system (~5 minutes). I pointed Claude Design at my design system repo. It spent about five minutes reviewing the tokens, components and patterns, then produced its own version. Its interpretation of the system, ready to build against.

2. Refining the system (~10 to 15 minutes). Claude's version wasn't quite right out of the box, but correcting it was fast. A mix of prompts ("use the condensed display face for headlines, not the body sans") and the direct editing tool. The editing tool is the thing that makes this feel like a design tool rather than a generator. You can reach in and fix things properly, not just re-roll the prompt.

3. Briefing the templates (loose, deliberately). I gave it a loose brief. Static and animated templates for the main social networks, covering simple posts, partnerships, product launches, blog content. I deliberately didn't over-specify. I wanted to see what it came back with.

4. Editing the output (~10 to 15 minutes). Same pattern as the system refinement. Prompts for conceptual changes, editing tool for pixel-level fixes. Content overflowed in a few of the denser templates, which is the usual tell that a system wasn't pressure-tested with real copy. Didn't take long to fix.

What worked

The design sensibility is genuinely good. Not generic AI-aesthetic good, actually considered. Hierarchy, rhythm, restraint in the right places. It understood my system rather than flattening it into something that felt like a template marketplace.

Export is the other quiet win. You can hand off to Claude Code, push to Canva, or download as Zip, HTML, PDF or PPTX. For a team that needs to move fast across channels and tools, that flexibility matters more than any single killer feature.

What didn't

• Content overflow on some templates. Minor, but a giveaway.
• Occasional bugginess. The usual tells of a product early in its life.
• No Figma integration yet. This will matter for a lot of people. Less so for me, because I'm building production work in code now, but I'd still use it for team review loops.

Would I replace Figma with it?

Not for production. Not yet. For final-mile design quality, the kind where you're fighting for the last 5% of polish that separates professional work from "pretty good", Figma (and increasingly, code) is still where that battle gets won.

But that's not really the question. The better question is: what's the cheapest way to go from concept to something the team can react to?

It's not that the tools make the old process faster. It's that they quietly make the old process the wrong one.

On that measure, Claude Design is the most useful tool I've used this year. An hour from "we need new social templates" to "here's a working system with a dozen variants across four networks" isn't a speed-up of my old workflow. It's a different workflow.

That's the thing I keep coming back to with AI-native design. It's not that the tools make the old process faster. It's that they quietly make the old process the wrong one.

I'll keep testing. But I went in expecting to hate it, and I don't. That's the headline.